Guardian Angels: LLM Personalization for Productivity and Security

What should AI look like from the perspective of one person trying to stay productive — and safe — as the internet fills with machine-generated deception? Gwern's answer is the "guardian angel": an AI trained not to be a generic assistant but to emulate you — your values, taste, and judgment — so that it can act for you at scale and screen what reaches you, the way a trusted family member screens an elderly relative's phone calls. Along the way he makes a bracing argument that today's chatbots are structurally misaligned with their users: the economic incentives run toward replacing you, not amplifying you.

Read it for the problem statement even more than the proposal. The opening sections — on scams, slop, and why "you are the bottleneck to be optimized away" — are the clearest articulation anywhere of what ordinary people will actually need from AI. The second half gets technical; the companion marks where general readers can stop with the full argument in hand.